The present invention relates to a finite impulse response filter, and particularly to such a filter in which a delay in a portion thereof has an adjustable or selectable delay period, and to an echo canceller and an Ethernet transceiver including such an FIR filter.
Finite impulse response (FIR) filters are extremely versatile digital signal processors that are used to shape and otherwise to filter an input signal so as to obtain an output signal with desired characteristics. FIR filters may be used in such diverse fields as Ethernet transceivers, read circuits for disk drives, ghost cancellation in broadcast and cable TV transmission, channel equalization for communication in magnetic recording, echo cancellation, estimation/prediction for speech processing, adaptive noise cancellation, etc. For example, see U.S. Pat. Nos. 5,535,150; 5,777,910; and 6,035,320, the contents of each of which are incorporated herein by reference. Reference is also made to the following publications: “An adaptive Multiple Echo Canceller for Slowly Time Varying Echo Paths,” by Yip and Etter, IEEE Transactions on Communications, October 1990; “Digital Signal Processing”, Alan V. Oppenheim, et al., pp. 155-163; AA 100 MHz Output Rate Analog-to-Digital Interface for PRML Magnetic-Disk Read Channels in 1.2 um CMOS@, Gregory T. Uehara and Paul R. Gray, ISSCC94/Session 17/Disk-Drive Electronics/ Paper FA 17.3, 1994 IEEE International Solid-State Circuits Conference, pp. 280-281; “72 Mb/S PRML Disk-Drive Channel Chip with an Analog Sampled Data Signal Processor”, Richard G. Yamasaki, et al., ISSCC94/Session 17/Disk-Drive Electronics/Paper FA 17.2, 1994 IEEE International Solid-State Circuits Conference, pp. 278, 279; “A Discrete-Time Analog Signal Processor for Disk Read Channels”, Ramon Gomez, et al., ISSCC 93/Session 13/Hard Disk and Tape Drives/Paper FA 13.1, 1993
ISSCC Slide Supplement, pp. 162, 163, 279, 280; and AA 50 MHz 70 mW 8 Tap Adaptive Equalizer/Viterbi Sequence Detector in 1.2 um CMOS@, Gregory T. Uehara, et al. 1994 IEEE Custom Integrated Circuits Conference, pp. 51 54, the contents of each being incorporated herein by reference.
Typically, an FIR filter is constructed in multiple stages, with each stage including an input, a multiplier for multiplication of the input signal by a coefficient, and a summer for summing the multiplication result with the output from an adjacent stage. The coefficients are selected by the designer so as to achieve the filtering and output characteristics desired in the output signal. These coefficients (or filter tap weights) are often varied, and can be determined from a least mean square (LMS) algorithm based on gradient optimization. The input signal is a discrete time sequence which may be analog or digital, while the output is also a discrete time sequence which is the convolution of the input sequence and the filter impulse response, as determined by the coefficients.
With such a construction, it can be shown mathematically and experimentally that virtually any linear system response can be modeled as an FIR response, as long as sufficient stages are provided. Because of this feature, and the high stability of FIR filters, such filters have found widespread popularity and are used extensively.
One problem inherent in FIR filters is that each stage requires a finite area on an integrated circuit chip. Additional area is required for access to an external pin so as to supply the multiplication or weighting coefficient for that stage. In some environments, the number of stages needed to provide desired output characteristics is large. For example, in Gigabit Ethernet applications it is preferred that every 8 meters of cable length be provided with 11 stages of FIR filter. In order to cover cable lengths as long as 160 meters, 220 FIR stages should be provided. In such environments, chip area on the integrated circuit is largely monopolized by the FIR stages.
Moreover, each FIR stage requires a finite amount of power and generates a corresponding amount of heat. Particularly where a large number of stages is needed, such power requirements become excessive and require significant mechanical adaptations to dissipate the heat.
The inventors herein have recently recognized that in some environments, not all stages of an FIR contribute significantly to the output. FIG. 1, for example, is a waveform showing signal amplitude versus time in an Ethernet echo cancellation application, where time (on the horizontal axis) is expressed in delay units for an FIR filter. The waveform shown in FIG. 1 represents an Ethernet transmission and its echo (or, reflection). As seen in FIG. 1, the waveform includes the near end echo at region 1, followed by a relatively quiet period in region 2, a relatively negligible signal at region 3, and the far end echo at region 4. One use of an FIR filter in such an Ethernet environment is to cancel the echo so as to distinguish more clearly between incoming signals and simple reflections of transmitted signals. However, the relatively negligible signal at region 3 contributes very little to the overall output of the FIR filter. The reason for this is that, whatever values of coefficients are set at the stages corresponding to region 3, those coefficients will be multiplied by a value which is approximately zero. Thus, contributions of those signals to the FIR output will be negligible, especially compared to regions 1, 2 or 4.
The inventors have considered simplifying the selection of coefficients by setting the coefficients corresponding to region 3 to zero, which would result in simpler algorithms needed to select coefficients. However, even with zeroed coefficients, the stages corresponding to region 3 still exist on the integrated circuit chip, stealing valuable surface area and power, and generating unwanted heat.